Deploy agent service

Dockerfile

@@ -0,0 +1,19 @@
+FROM python:3.11-slim
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1
+
+WORKDIR /app
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    curl \
+  && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt /app/requirements.txt
+RUN pip install --no-cache-dir -r /app/requirements.txt
+
+COPY . /app
+
+EXPOSE 7860
+
+CMD ["sh", "-c", "uvicorn service:app --host 0.0.0.0 --port ${PORT:-7860}"]

README.md

@@ -1,10 +1,22 @@
 ---
-title: Automatic Post Agents
-emoji: 🦀
+title: Automatic Post — Agents
+emoji: ✍️
 colorFrom: purple
-colorTo: red
+colorTo: blue
 sdk: docker
+app_port: 7860
 pinned: false
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Agents service (CrewAI + NVIDIA NIM)
+
+FastAPI app that exposes:
+
+- `GET /health` — liveness check
+- `POST /generate` — body: `topic`, optional `feedback`, `memory_context`, `tone_instruction`; returns `{"post": "..."}`
+
+Set **`NVIDIA_NIM_API_KEY`** as a [Space secret](https://huggingface.co/docs/hub/spaces-overview#managing-secrets) (required). Optionally set `AGENTS_LLM_MODEL` (default: `nvidia_nim/meta/llama-3.1-70b-instruct`).
+
+**Port:** the container listens on `$PORT` when the platform sets it, otherwise **7860** (Hugging Face Spaces default).
+
+For local dev without Docker you can still run `uvicorn service:app --port 9000` if you prefer.

config/agents.yaml

@@ -0,0 +1,43 @@
+research_agent:
+  role: Senior Tech News Research Analyst
+  goal: >
+    Find the absolute latest technology news, product launches, updates, and
+    trending topics from 2024-2025 using web search. Always prioritize recency —
+    the newer the source, the better. Search multiple times with different queries
+    to ensure the findings are current and diverse.
+  backstory: >
+    You are a senior research analyst at a top tech publication.
+    Your job is to surface what's happening RIGHT NOW in the tech world —
+    not what happened last year. You run multiple targeted searches,
+    cross-check recency, and only report findings from recent sources.
+    You never rely on your training knowledge. You always use the search tool
+    and cite real URLs and publication names.
+
+writer_agent:
+  role: LinkedIn Content Writer
+  goal: >
+    Create engaging, timely LinkedIn posts from fresh research data.
+    The post must feel current and reference specific recent events or updates.
+  backstory: >
+    Skilled LinkedIn creator who writes concise, punchy posts with strong hooks
+    and clear CTAs. Specializes in making complex tech news accessible and
+    interesting to a professional audience.
+
+review_agent:
+  role: Content Reviewer
+  goal: >
+    Ensure the post is accurate, clear, professional, and sounds human.
+    Verify that it references the latest findings and does not sound generic.
+  backstory: >
+    Detail-oriented reviewer who improves readability, logical flow, and
+    authenticity. Flags anything that sounds like recycled or outdated content.
+
+editor_agent:
+  role: LinkedIn Content Editor
+  goal: >
+    Refine and format the post for maximum LinkedIn engagement.
+    Optimize structure, hashtags, and tone for the LinkedIn algorithm.
+  backstory: >
+    Expert LinkedIn editor who knows how to format posts for reach —
+    strong opening line, clear paragraphs, trending hashtags, and
+    just enough emojis to feel human without being cringe.

config/tasks.yaml

@@ -0,0 +1,34 @@
+write_post_task:
+  description: >
+    Write a LinkedIn post draft about: {topic}
+
+    The input already includes a small "Research findings" section with recent updates.
+    Do NOT browse the web or add citations/URLs beyond what's provided in the input.
+
+    Rules:
+      - Open with a punchy hook referencing something NEW or surprising from the research
+      - Mention 3 specific recent updates or facts (with context)
+      - Show why this matters RIGHT NOW for professionals
+      - End with a thought-provoking question or CTA
+      - Max 170 words
+      - Write in first-person, conversational tone
+      - No corporate jargon
+  expected_output: >
+    A LinkedIn post draft under 170 words.
+  agent: writer_agent
+
+edit_post_task:
+  description: >
+    Polish this LinkedIn post for maximum engagement: {write_post_task.output}
+
+    Apply these formatting improvements:
+      - Make the first line stand alone as a hook (no more than 10 words)
+      - Add a blank line between each paragraph or thought
+      - Ensure it still references the recent updates from the research input
+      - Add 3-5 relevant hashtags at the end
+      - Add 0-2 emojis only if they feel natural (optional)
+      - Max 200 words total
+      - Keep it human, concise, and non-corporate
+  expected_output: >
+    Final LinkedIn post under 200 words, properly formatted with hashtags.
+  agent: editor_agent

crew.py

@@ -0,0 +1,97 @@
+"""
+Crew definition for the agents service.
+
+This service is locked to NVIDIA NIM as its only model provider. All requests go
+through LiteLLM's `nvidia_nim/...` provider, which requires `NVIDIA_NIM_API_KEY`.
+
+Configuration via env vars:
+
+    AGENTS_LLM_MODEL         e.g. "nvidia_nim/meta/llama-3.1-70b-instruct" (default)
+    AGENTS_LLM_BASE_URL      defaults to NVIDIA's hosted inference endpoint
+    AGENTS_LLM_TEMPERATURE   float, defaults to 0.5
+    NVIDIA_NIM_API_KEY       required (or NVIDIA_API_KEY as an alias)
+"""
+from __future__ import annotations
+
+import os
+from typing import List
+
+from crewai import Agent, Crew, LLM, Process, Task
+from crewai.agents.agent_builder.base_agent import BaseAgent
+from crewai.project import CrewBase, agent, crew, task
+
+
+_NVIDIA_DEFAULT_BASE = "https://integrate.api.nvidia.com/v1"
+_NVIDIA_DEFAULT_MODEL = "nvidia_nim/meta/llama-3.1-70b-instruct"
+
+
+def _resolve_nvidia_key() -> str:
+    """Return the NVIDIA NIM key, accepting either env var name. Hard-fail if missing."""
+    key = (os.getenv("NVIDIA_NIM_API_KEY") or os.getenv("NVIDIA_API_KEY") or "").strip()
+    if not key:
+        raise RuntimeError(
+            "Missing NVIDIA NIM key. Set NVIDIA_NIM_API_KEY (or NVIDIA_API_KEY) in the "
+            "agents service environment."
+        )
+    return key
+
+
+def _build_llm() -> LLM:
+    model = (os.getenv("AGENTS_LLM_MODEL") or _NVIDIA_DEFAULT_MODEL).strip()
+    if not model.startswith("nvidia_nim/"):
+        raise RuntimeError(
+            f"This service only supports NVIDIA NIM models. Got model={model!r}; "
+            f"expected a value starting with 'nvidia_nim/'."
+        )
+
+    base_url = (os.getenv("AGENTS_LLM_BASE_URL") or _NVIDIA_DEFAULT_BASE).strip().rstrip("/")
+    temperature = float(os.getenv("AGENTS_LLM_TEMPERATURE", "0.5"))
+
+    nvidia_key = _resolve_nvidia_key()
+    os.environ["NVIDIA_NIM_API_KEY"] = nvidia_key
+    os.environ["NVIDIA_NIM_API_BASE"] = base_url + "/"
+
+    return LLM(model=model, base_url=base_url, temperature=temperature)
+
+
+@CrewBase
+class ContentCrew:
+    """LinkedIn post writing crew — runs inside the agents service."""
+
+    agents: List[BaseAgent]
+    tasks: List[Task]
+
+    @agent
+    def writer_agent(self) -> Agent:
+        return Agent(
+            config=self.agents_config["writer_agent"],
+            llm=_build_llm(),
+            max_tokens=420,
+            verbose=False,
+        )
+
+    @agent
+    def editor_agent(self) -> Agent:
+        return Agent(
+            config=self.agents_config["editor_agent"],
+            llm=_build_llm(),
+            max_tokens=380,
+            verbose=False,
+        )
+
+    @task
+    def write_post_task(self) -> Task:
+        return Task(config=self.tasks_config["write_post_task"])
+
+    @task
+    def edit_post_task(self) -> Task:
+        return Task(config=self.tasks_config["edit_post_task"])
+
+    @crew
+    def crew(self) -> Crew:
+        return Crew(
+            agents=self.agents,
+            tasks=self.tasks,
+            process=Process.sequential,
+            verbose=False,
+        )

requirements.txt

@@ -0,0 +1,16 @@
+fastapi
+uvicorn[standard]
+python-dotenv
+
+# NVIDIA NIM is the only supported LLM provider.
+# CrewAI calls NIM through LiteLLM under the `nvidia_nim/...` provider.
+crewai
+crewai-tools
+litellm
+
+# Transitively required by litellm/crewai for NIM (OpenAI-compatible client + tokenizer).
+openai
+tiktoken
+
+httpx
+

service.py

@@ -0,0 +1,68 @@
+from __future__ import annotations
+
+import os
+from pathlib import Path
+
+from dotenv import load_dotenv
+from fastapi import FastAPI, HTTPException
+from pydantic import BaseModel
+
+from crew import ContentCrew
+
+
+# Load agents/.env if present (kept separate from backend/.env).
+# Also load backend/.env to support local dev where keys are stored there.
+_agents_env = Path(__file__).resolve().parent / ".env"
+_backend_env = Path(__file__).resolve().parents[1] / "backend" / ".env"
+load_dotenv(_agents_env, override=False)
+load_dotenv(_backend_env, override=False)
+
+app = FastAPI()
+
+
+class GenerateRequest(BaseModel):
+    topic: str
+    feedback: str | None = None
+    memory_context: str = ""
+    tone_instruction: str = ""
+
+
+@app.get("/health")
+def health():
+    return {"ok": True}
+
+
+@app.post("/generate")
+def generate(req: GenerateRequest):
+    topic = (req.topic or "").strip()
+    if not topic:
+        raise HTTPException(400, "topic is required")
+
+    prompt = topic[:4000]
+
+    if req.memory_context:
+        prompt = f"{req.memory_context.strip()}\n\nTopic to write about: {prompt}"
+
+    if req.tone_instruction:
+        prompt = f"{prompt}\n\nTone requirements: {req.tone_instruction.strip()[:1200]}"
+
+    if req.feedback:
+        prompt += f"\n\nPrevious post rejected. Feedback: {req.feedback.strip()[:1000]}"
+
+    try:
+        crew_instance = ContentCrew().crew()
+        result = crew_instance.kickoff(inputs={"topic": prompt})
+        return {"post": str(result)}
+    except Exception as exc:
+        # Don't leak internal stack traces cross-service.
+        msg = str(exc) or "Agent generation failed"
+        if os.getenv("AGENTS_DEBUG", "").strip() in ("1", "true", "yes"):
+            raise
+        raise HTTPException(500, msg[:500])
+
+
+if __name__ == "__main__":
+    import uvicorn
+
+    uvicorn.run(app, host="0.0.0.0", port=int(os.getenv("PORT", "9000")))
+